Getting cited in ChatGPT is not a commercial win. Your page can supply the fact while a competitor gets the recommendation.

That is the distinction I care about when I read an AI visibility report. A citation proves the engine used your source, not that it put your business on the shortlist. In one dataset, 62% of citations had no brand mention at all, and only 13.2% were both cited and mentioned. Google AI Overviews show the same split: they cite brand sources in 87% of answers but name a brand in only 20.7%. A report can celebrate the first number while the business feels the second.

If you are shopping for help getting recommended across ChatGPT, Gemini and Perplexity, do not buy citation volume. Buy the work that makes a recommendation defensible. Judge the provider on relevant recommendation coverage and attributable enquiries, not how often an engine borrows your homework.

Your source can make the answer while your business misses the shortlist

I separate three outcomes that visibility reports keep bundling together:

  • Citation: the engine uses your page as a source.
  • Mention: the engine names your business.
  • Recommendation: the engine tells the buyer to contact you, with a reason.

They are not interchangeable. What analysts call a ghost citation is the trap: your page gets used for a fact while your brand stays invisible. In cited datasets, 48.73% of ChatGPT citations came from third parties, rising to 82% off-site in recommendation-style queries. Your own page is only part of the source set a buyer’s answer draws on.

The recommendation problem is bigger than publishing one helpful page. The brands that get picked appear across the sources the engine uses. One analysis found that at least 92.7% of brands recommended by ChatGPT appear in cited content, while a brand appearing in six or more cited URLs is six times more likely to be recommended. In the same analysis, 96% of brands appearing for the first time on a cited page are not recommended.

Position within those sources matters too: top-three source positions get the brand named 96.1–98.8% of the time, versus about 21.6% at position eight or lower.

Those findings are not a placement guarantee. They explain why I do not accept “we got your page cited” as the finish line. Being useful to the answer is not the same as being chosen by it.

Illustration of a cited page feeding a competitor’s shortlist

Stop counting homework questions as purchase orders

I use the same filter here that I use in paid search: what is the person actually trying to do?

An informational prompt like what is waterproofing asks for an explanation. A supplier-selection prompt like who should I hire for basement waterproofing in Charlotte under $15k with a transferable warranty asks for a shortlist.

If a provider blends both into one visibility score, it counts homework questions as purchase orders. Say you run a home services business spending $20k a month on ads. Ten citations for how questions are not ten recommendations to hire you. The distinction matters before anyone starts claiming commercial progress.

The answers also move. Across one 43,000-keyword month, AI Overview content had a 70% chance of differing between checks, lasted 2.15 days on average, and had 45.5% new cited URLs each time. One screenshot cannot tell me whether a business consistently makes the shortlist.

Before I accept a visibility count, I want the test defined:

  1. Fix the buying prompt, including the market, service and constraints.
  2. Name the engines and search surfaces being tested.
  3. Repeat the checks, recording date, location and session.
  4. Keep informational citations separate from commercial recommendations.

Then I want mention, citation, link and accuracy reported separately, with the raw log behind them. Recommendation coverage needs its own view, not a footnote under a blended score.

Measure the questions buyers use to choose a supplier. Otherwise, a provider can improve the score without improving your place in the decision.

Buy the work that makes a recommendation defensible

A recommendation is a claim the engine has to support with sources. I want the business to give it clear facts, then earn corroboration beyond its own website.

Start with an answer-first service page for each service and city you actually serve. Put the information a buyer needs where it can be found:

  • Price ranges or plans.
  • Eligibility rules.
  • Turnaround times.
  • Warranty terms.

Use plain numbers where they exist. Mark up the relevant facts with Organization, Service and FAQ schema. Handle the unglamorous access work too: allow OAI-SearchBot in robots.txt, keep pages indexed and snippet-eligible for AI Overviews, and keep Google Business Profile details consistent.

But do not confuse making a page readable with making the business recommendable. In a test of 112 startups, own-page formatting score had no correlation with being recommended. What the deck calls optimization can still amount to paperwork the engine can quote.

The same test found Reddit mentions had the strongest correlation with recommendations at r=+0.40, followed by referring domains at +0.32, while own-page tweaks sat at -0.10. A 75,000-brand study points in the same direction: web mentions correlated at 0.664 versus 0.218 for backlinks.

That is why I look beyond another informational post. I want evidence where buyers compare businesses: review profiles, directories, honest listicles with 10 to 20 options, YouTube walkthroughs and relevant Reddit threads the business did not manufacture.

The source mix differs by engine too. The cited analysis describes ChatGPT leaning toward consensus, Gemini toward official sites at 52.15%, and Perplexity toward niche and Reddit sources near 24%. One engine screenshot does not settle the question across all three.

The cost of this alternative is real work outside a content calendar. Clear service pages, technical fixes and earned corroboration are not the same deliverable as another batch of articles. If a proposal adds pages but no outside proof, it leaves the recommendation problem unfinished.

Keep traditional SEO as the floor, not a rival project

None of this excuses broken basics. If a service page is blocked, buried three clicks deep or replaced every quarter without redirects, the engine has no stable page to use.

I treat classic SEO as the floor:

  • Keep service pages crawlable and indexed.
  • Tie titles and headings to the service and city the business actually serves.
  • Point internal links at the pages that sell the work.
  • Keep first-party performance reporting alongside the repeated answer checks.

The baseline checklist belongs in the proposal before the clever visibility score does. I also keep a glossary separating mention, citation and recommendation open when I read vendor reports. Terminology gets suspiciously flexible when a target is missed.

Fix access and service-page clarity before paying anyone to chase mentions. A new label for search does not remove the old plumbing.

Demand a report that separates visibility from revenue

A useful report reads like a lab notebook, not a trophy cabinet. I want enough detail to see what happened, repeat the check and understand what the result does not prove.

That means three distinct layers:

  1. The test. Exact buyer prompts, selected engines and surfaces, runs per query, dates, locations and sessions.
  2. The answer. Mentions, citations, links and accuracy reported separately, plus whether the business actually makes the recommendation set.
  3. The commercial trail. Attributable enquiries where a caller, form or chat names the assistant, or where its referral identifies the visit.

The reporting baseline needs the raw log attached. If the provider tests ChatGPT search, AI Overviews, AI Mode, Perplexity, Gemini and Copilot, I want to see the results for each, not a percentage that hides the misses.

Then I want the provider to stop attributing when the evidence stops. An enquiry that names the assistant is useful evidence. An enquiry that merely arrives during a month when citations rise is not the same thing.

I use a method for measuring AI visibility without tracking fake rankings because a blended score tells me nothing about which prompt produces a shortlist and which produces homework. Nor does it tell me whether anybody gets in touch.

Recommendation coverage and attributable enquiries belong beside each other, not collapsed into one claim. If the report cannot show the misses by prompt and engine, it cannot explain the wins.

Shortlist clipboard showing three selected businesses beside unused cited pages

Choose the provider that ships fixes and shows the misses

When I compare companies offering help with ChatGPT recommendations or visibility across AI search, I start with one question: who actually ships the work?

Much of what gets sold as AI search help measures the problem rather than fixing it. The listed provider ranges illustrate how different those purchases can be:

  • Platforms at $29–$999 per month count mentions but do not publish fixes.
  • Audits at $99–$5,000 list problems and leave implementation on the buyer’s desk.
  • Retainers at $497–$12,000 per month vary in what they implement, with examples also carrying $200–$500 in tool fees and onboarding near $1,750.

The price is not the scope. I want the proposal to name the pages the provider will publish, the technical fixes it will ship itself, the third-party coverage it will earn honestly and the prompts it will test weekly. Leave the misses in the log.

Anything promising placement without that work list is selling a scoreboard as labor.

That honesty extends to guarantees. OpenAI says ChatGPT search placement is not guaranteed; Google says no one can guarantee top ranking; Bing says optimization does not guarantee grounding or citations. I want a provider that says no to guarantee questions, publishes its method and prices, and treats self-ranked best-agency lists and secret scores as sales material rather than proof.

The shortcuts deserve the same scrutiny. Reddit seeding, paid links without proper tags and fabricated reviews do not become sound strategy because someone puts AI visibility on the invoice. They run into Reddit spam rules, Google link spam and scaled content policies, and the FTC fake review rule. I am not paying for a mention that comes with a trust problem attached.

The filter stays simple: published pages, shipped fixes, earned corroboration and an honest record of where the business still loses.

That is where groas fits for me. It runs paid and organic search as one system instead of two vendors arguing about credit. Specialized models handle continuous execution; a named human strategist owns direction and accountability. Actions are logged so I can inspect what changes rather than take a polished report on faith.

The live trial and flat monthly fee also put a different proposition on the table from percentage-of-spend billing that rewards a bigger budget. I want execution I can inspect, not more hours to fund or another dashboard to manage. The point is not a guaranteed citation. It is doing the unglamorous work this argument demands: clear service pages, clean tracking and corroboration where buyers compare.

Keep paying for citation volume and the consequence is predictable. Next quarter’s report can get thicker and its score can climb while the enquiry line stays flat. The buyer asks ChatGPT who to hire and calls the business the answer actually recommends.

Your page supplies the facts. Your competitor gets the call. Fix the recommendation gap or keep funding someone else’s shortlist.